Snowflake CoWork Defined
Snowflake CoWork is the personal work agent Snowflake announced at Summit 2026, previously known as Snowflake Intelligence. Where CoCo answers questions, CoWork takes action. It reasons across tasks, automates routine work, and moves a business question through to a decision and an action without the user having to assemble the pieces by hand.
An autonomous work agent that reasons across Snowflake data, automates multi-step tasks, and operates with enterprise context governed by Horizon Context.
A general-purpose copilot. CoWork is grounded in Snowflake's governance and semantic layer rather than in open-web context, which is what makes it useful for enterprise workflows and what limits it to surfaces Snowflake can reach.
Knowledge workers in finance, sales, operations, and analytics who need an agent that understands both their data and the metric definitions their team has approved.
How CoWork Works
A CoWork task moves through five steps:
- 1
The user describes the work in natural language inside Snowflake — for example "build a Q3 forecast variance for North America and notify the regional VPs."
- 2
CoWork plans the steps needed to complete the task, consulting Horizon Context for the metric definitions involved.
- 3
CoWork executes each step querying Snowflake Semantic Views, running the right calculations, and pulling in supporting context as needed.
- 4
CoWork applies governance at every step with Snowflake RBAC and row-level masking enforced as the agent moves through the data.
- 5
CoWork returns a result and a record including the actions taken, the definitions used, and the data the agent saw on the way through.
According to Snowflake's CoWork announcement, the combined CoCo and CoWork experience hits 83% accuracy on complex enterprise queries when paired with Cortex Sense, compared with 47% without that enrichment layer.
What CoWork Solves
CoWork is built for the work that doesn't fit cleanly into a single question and a single answer.
Multi-step workflows the business already runs. Forecast variance, churn analysis, account reviews, monthly close support. CoWork sequences the steps automatically and applies governed definitions at each one.
Routine analytics tasks that drain analyst time. Repetitive data pulls, dashboard refreshes, anomaly checks. CoWork handles the steps and surfaces the result.
Agent-driven actions, not just agent-driven answers. CoWork can write results back into Snowflake, post notifications, and trigger downstream workflows, which moves agentic AI from advisory to operational.
Where CoWork Stops
CoWork operates inside Snowflake. It governs the steps it runs natively, but the rest of the enterprise data stack lives outside the Snowflake perimeter. The CFO opens a Power BI dashboard. The finance team reconciles in Excel. The product team reviews customer activity in Tableau. The marketing analyst asks a Claude agent for last quarter's campaign performance.
If those tools and agents don't read the same Snowflake Semantic Views that CoWork is using, the result is metric drift between what CoWork did and what the business sees afterwards. A CoWork workflow can book a forecast adjustment that the Excel pivot the CFO opens an hour later reports as a different number, because the definition the agent used and the definition Excel computed don't match.
How to Extend CoWork's Definitions to Every Tool the Business Uses
A universal semantic layer like AtScale for Snowflake reads the same Snowflake Semantic Views CoWork relies on and serves those same governed definitions to Power BI, Excel, Tableau, Looker, Google Sheets, and external AI agents through standard interfaces (XMLA, MDX, DAX, JDBC, MCP). The action CoWork takes inside Snowflake and the number a business user sees in their tool of choice come from the same source.
CoWork governs inside Snowflake. AtScale carries those same definitions to Power BI, Excel, Tableau, and every AI agent the enterprise runs.
What the Anthropic Benchmark Says About CoWork's Accuracy Ceiling
CoWork's accuracy is bounded by the semantic layer feeding it. In How Anthropic's AI Accuracy Went from 21% to 95%, AtScale CTO Dave Mariani lays out two benchmarks worth working through. Claude, given raw access to thousands of SQL files, dashboards, transforms, and notebooks at Anthropic, answered correctly about 21% of the time. After a semantic layer and a "check the semantic layer first" rule went in, accuracy climbed to roughly 95%, with some domains landing close to 99%.
AtScale's production benchmark with the commercial banking arm of a Tier 1 global bank cut compute by up to 21,000x and lifted accuracy from around 70% to 100% on a set of five common questions, with projected savings of roughly $9 million a year on that one question set. Horizon Context gets CoWork to a strong baseline inside the data cloud, and the same accuracy and cost profile only carries out to the tools and agents the business runs on if a universal semantic layer extends those definitions beyond the Snowflake perimeter.
After adding a semantic layer to Claude's analytics workflow.
Accuracy from ~70% to 100% on five common questions, $9M/yr projected savings.
A Skill Is Not a Semantic Layer
A common shortcut is to write a skill file telling CoWork (or any agent) how the business defines its metrics. Skills work, and Anthropic got real lift from theirs, but the AtScale post Semantic Layer vs. Skill for AI Agents makes the limit clear: "A skill describes your business. It can't enforce it."
A skill is a static instruction, and a model treats it as a suggestion rather than a contract. A semantic layer is a running engine between the agent and the warehouse that enforces the definition, governs access, and routes every query to the cheapest correct path.
Frequently Asked Questions
Is Snowflake CoWork the same as Snowflake Intelligence?
Yes. Snowflake Intelligence was rebranded as CoWork at Summit 2026. The product is the same lineage, with continued investment in agentic workflows on top of Horizon Context.
What's the difference between CoCo and CoWork?
CoCo answers questions. CoWork takes action across multi-step tasks. Both consume Horizon Context for governed business meaning.
Does CoWork work outside Snowflake?
CoWork executes inside Snowflake's environment. For the actions CoWork takes to align with what the business sees in Power BI, Excel, Tableau, or external agents, the same definitions need to be served by a universal semantic layer that reaches those tools.
Can CoWork write data back to Snowflake?
Yes. CoWork is built for multi-step agentic workflows that include actions, not just answers, which is a meaningful step beyond the read-only behavior of most BI copilots.
What governance applies to CoWork actions?
CoWork inherits Snowflake's role-based access control, row-level masking, and Horizon Context governance, which means actions and answers stay inside the data team's approved rules.
Related Terms
Key Takeaways
Snowflake CoWork is the personal work agent for knowledge workers, previously known as Snowflake Intelligence, built on Horizon Context for governed business meaning.
CoWork takes action across multi-step workflows, where CoCo answers single questions, and both share the same semantic foundation inside Snowflake.
The accuracy and cost profile of CoWork only extends to Power BI, Excel, Tableau, and external agents when a universal semantic layer serves the same governed definitions to those tools.